[python] Fix perf. hit for large enums on ingest from AnnData

[johnkerl]

Scenario:

• User has an AnnData obs with a string-valued column with a large number of distinct values (i.e. there is no reason to have this be categorical)
• They do AnnData adata.write_h5ad(...)( and later adata = anndata.read_h5ad(...)
• AnnData has force-converted this to a categorical within the read-back adata.obs Pandas DataFrame -- before we ever see it
• Then, at tiledbsoma.io.from_anndata, we are too trusting: given a categorical column in AnnData adata.obs, we make a categorical column in TileDB-SOMA
• For cases like 4M distinct values in a categorical column, this faithful/trusting approach is a performance negative

Solution:

• The tiledbsoma.io ingest code should inspect each categorical column, and if the number of categories exceeds some threshold (212 and 215-1 have both been proposed), store it as a simple primitive typed column, i.e, with no enumerations.
• One negative is that outgesting back to AnnData would produce a string column, while the input had categorical. On discussion we believe:
  • Cases of huge enumerations are more likely to be mistakes than intentional
  • If we were to keep a metadata 'outgest hint' for tiledbsoma dataframe string columns, saying, convert back to categorical on outgest -- this would be prohibitive if the input were ordered. And if we can't do it for ordered, we shouldn't do it for non-ordred (that would be confusing).

[sc-59407] [sc-59595]